CN-117706623-B - Microseism speed inversion method based on deep learning and wave equation combined driving

CN117706623BCN 117706623 BCN117706623 BCN 117706623BCN-117706623-B

Abstract

The invention relates to a deep learning and wave equation combined driving microseism speed inversion method, which comprises the steps of dispersing an underground model, setting detectors, establishing an underground speed model, calculating waveform information of a random seismic source at the position of each detector by utilizing a finite difference algorithm, inputting a waveform record, a speed model and source coordinates into Unet neural networks in batches, obtaining a preliminary speed structure by pure data driving calculation, obtaining a wave equation by using a finite difference solution, obtaining an earthquake record, calculating structural similarity between a predicted speed and a real speed, setting an average value of the structural similarity as dynamic weight of a physical driving loss value, extracting a mapping relation between waveform characteristics and the speed model by utilizing the neural networks, and inverting and predicting the underground speed structure. The invention realizes the artificial intelligent real-time prediction of the microseism velocity model, and the velocity model obtained by inversion has high precision and strong interpretability.

Inventors

ZENG XIAOBAO
LI LEI
PAN XINPENG
LIU JIANXIN

Assignees

中南大学

Dates

Publication Date: 20260508
Application Date: 20231214

Claims (4)

1. The deep learning and wave equation combined driving microseism speed inversion method is characterized by comprising the following steps of: S1, dispersing an underground model by using grids, and setting n detectors in the underground model, wherein n is a natural number; s2, establishing an underground speed model, setting a region where a random seismic source is located, and calculating waveform information of the random seismic source at each detector position by using a finite difference algorithm; s3, inputting Unet the waveform record, the velocity model and the seismic source coordinates into a neural network in batches after corresponding; s4, performing pure data driving calculation to obtain a preliminary speed structure, performing physical driving calculation, and solving a wave equation by using finite difference to obtain a simulated seismic waveform record; s5, the loss function consists of a speed error calculated by pure data driving and a waveform error calculated by physical driving; S6, calculating the structural similarity between the predicted speed and the real speed, and setting the average value of the structural similarity as the dynamic weight of the physical driving loss value; And S7, extracting waveform characteristics by utilizing a neural network, and inversely predicting an underground speed structure.
2. The method for deep learning and wave equation joint driving microseismic velocity inversion of claim 1, wherein in S5, the loss function is represented by the following formula: Where L d is a data driven loss function, x i is the true speed value, N is the sum of pixel points of a single Zhang Sudu graph for predicting the speed value; where L p is the loss function of the physical drive, j is the number of sampling points, r is the number of detectors, As the waveform data to be input, Waveform data synthesized for a physical driving process; wherein L is a loss function of the whole neural network, when epoch < epochs _syn, pure data driving is performed, and when epoch is greater than or equal to epochs _syn, physical driving is performed, wherein w1 and w2 are dynamic weights.
3. The method for deep learning and wave equation joint driving microseismic velocity inversion according to claim 1, wherein the dynamic weight calculation formula in S6 is: w1+w2=1 Wherein SSIM is structural similarity, SSIM P is structural similarity of P-wave velocity, SSIM S is structural similarity of S-wave velocity, and Batchsize is bulk size.
4. The method for inversion of deep learning and wave equation jointly driven microseismic velocity according to claim 1, wherein the neural network extracts waveform features to invert the velocity model in S7 is as follows: v=Net(d;θ) Wherein v≡v p ,v s is a predicted velocity value, net is Unet neural network model, d is microseism waveform data, and θ is neural network weight.

Description

Microseism speed inversion method based on deep learning and wave equation combined driving Technical Field The invention relates to the technical field of geophysical microseism monitoring, in particular to a microseism speed inversion method driven by combination of deep learning and wave equation. Background Hydraulic fracturing is one of the key technologies for economic development of unconventional oil and gas reservoirs at present. Microseism monitoring technology is one of the most effective methods for monitoring the effect of hydraulic fracturing in real time. By monitoring weak earthquakes induced by engineering operations such as hydraulic fracturing and the like, the space size, the extension form and the like of the fracturing cracks can be revealed, and the construction design and the effect of the fracturing are facilitated to be optimized. To ensure safe and efficient reservoir development, accurate imaging of the fracture is required. The focus positioning is directly related to the crack imaging effect, and the speed is taken as an indispensable parameter in the micro-seismic data processing process, so that the positioning precision is directly affected. The traditional inversion method of the microseism velocity model is mainly travel time inversion and full waveform inversion, but is limited by the low signal-to-noise ratio of actual data and the accuracy of an initial velocity model, so that an inversion result is often trapped in a local minimum value, and the inversion error is larger and the efficiency is not high. In recent years, the development of deep learning brings new prospects to the geophysical field, especially the seismic exploration field. The deep learning neural network can be used for effectively extracting the potential mapping relation between the seismic data and the velocity model, but the pure data driving deep learning method depends on the quality of the data, and no physical information is introduced, so that the generalization capability is not strong and the interpretability is poor. Disclosure of Invention Aiming at the problem that the conventional travel time and full waveform speed inversion method can be influenced by noise and initial speed in the microseism data processing, the invention provides the microseism speed inversion method driven by the combination of the deep learning and the wave equation, and the precision of microseism speed inversion is improved by the nonlinear fitting capacity of a neural network and the physical information of introducing the wave equation constraint, so that the generalization and the interpretability of the network are improved, and a reliable speed model is provided for the seismic source positioning. In order to achieve the above object, the present invention is realized by the following technical scheme: a deep learning and wave equation combined driving microseism speed inversion method comprises the following steps: S1, dispersing an underground model by using grids, and setting n detectors in the underground model, wherein n is a natural number; s2, establishing an underground speed model, setting a region where a random seismic source is located, and calculating waveform information of the random seismic source at each detector position by using a finite difference algorithm; s3, inputting Unet the waveform record, the velocity model and the seismic source coordinates into a neural network in batches after corresponding; s4, performing pure data driving calculation to obtain a preliminary speed structure, performing physical driving calculation, and solving a wave equation by using finite difference to obtain a simulated seismic waveform record; s5, the loss function consists of a speed error calculated by pure data driving and a waveform error calculated by physical driving; The loss function is as follows: Where L d is a data driven loss function, x i is the true speed value, N is the sum of pixel points of a single Zhang Sudu graph for predicting the speed value; where L p is the loss function of the physical drive, j is the number of sampling points, r is the number of detectors, As the waveform data to be input,Waveform data synthesized for a physical driving process; wherein L is a loss function of the whole neural network, when epoch < epochs _syn, pure data driving is performed, and when epoch is greater than or equal to epochs _syn, physical driving is performed, wherein w1 and w2 are dynamic weights. S6, calculating the structural similarity between the predicted speed and the real speed, and setting the average value of the structural similarity as the dynamic weight of the physical driving loss value; The calculation formula of the dynamic weight is as follows: w1+w2=1 Wherein SSIM is structural similarity, SSIM P is structural similarity of P-wave velocity, SSIM S is structural similarity of S-wave velocity, and Batchsize is bulk size. And S7, extracting waveform characteristics by utilizing a